An agent only uses a tool when it thinks it should. Custom instructions are how you tilt that decision your way.
Before you write instructions
You need a project with at least one table, an API key, and your agent connected via MCP. Once those are in place, come back here.
#The recipe
- Name the trigger — what topic, command or word should make the agent reach for the data.
- Name the connector — match the label you used when adding the MCP server.
- Name the action — read first? write after? both?
#Examples
text
# Fitness
When I talk about workouts, use my "Webenta Index — fitness" connector.
Read the last 7 days of sessions before answering. After any completed workout,
insert a row into "sessions" with date, exercise, sets, reps and notes.text
# Language learning
When I ask about my vocabulary or study sessions, use "Webenta Index — languages".
Read my recent reviews and study time before answering. After a study session,
insert a row into "study_sessions" with language, duration and quality rating.text
# Journal
If I mention "today" or "yesterday", append an entry to the "journal" table
in Webenta Index. Read the last 5 entries first so you have context.Per-project instructions beat global ones
Most runtimes (Claude Projects, ChatGPT Projects, custom Gems, Cursor Rules) let you scope instructions to a workspace. Use those — global instructions get crowded fast.